Kubernetes v1.36 introduces an architectural separation between the Workload API, acting as a static template, and the new PodGroup API, which manages runtime state for batch scheduling. This decoupling enables atomic workload processing, enhances scalability through per-replica status sharding, and integrates advanced features like topology-aware scheduling and Dynamic Resource Allocation support.
Key points
The Workload API now serves exclusively as a static template, while the PodGroup API handles runtime scheduling state.
Decoupling the APIs allows the kube-scheduler to process PodGroup information directly, improving performance and scalability.
The new architecture enables atomic gang scheduling and workload-aware preemption for complex AI/ML batch jobs.
ResourceClaim support in the new API facilitates integration with Dynamic Resource Allocation (DRA) for coordinated resource management.